import json
import pandas as pd
import difflib
import base64
from IPython.display import HTML, display

print("✅ Loaded dependencies")

✅ Loaded dependencies

# OPTIONAL: Load old eval results (for historical reference)
# You can skip this if you only want to do the audit batch analysis below
RESULTS_FILE = "./final_evaluation_results.json"

try:
    with open(RESULTS_FILE, 'r') as f:
        results = json.load(f)
    print(f"✅ Loaded {len(results)} old eval results")
except FileNotFoundError:
    print("ℹ️  Old eval results not found (this is optional)")
    print("   Skip to 'Label Noise Audit' section below for the main analysis")
    results = []

ℹ️  Old eval results not found (this is optional)
   Skip to 'Label Noise Audit' section below for the main analysis

def highlight_differences(truth, pred):
    """
    Compares two strings word-by-word and highlights differences.
    Returns tuple: (HTML_Ground_Truth, HTML_Prediction)
    """
    # Split into words for comparison
    a_words = truth.split()
    b_words = pred.split()
    
    # Use SequenceMatcher to find the differences
    matcher = difflib.SequenceMatcher(None, a_words, b_words)
    
    html_truth = []
    html_pred = []
    
    for opcode, a0, a1, b0, b1 in matcher.get_opcodes():
        # EQUAL: Text matches, just append it
        if opcode == 'equal':
            html_truth.append(" ".join(a_words[a0:a1]))
            html_pred.append(" ".join(b_words[b0:b1]))
            
        # INSERT: Model added words (Green in Pred)
        elif opcode == 'insert':
            inserted_text = " ".join(b_words[b0:b1])
            html_pred.append(f'<span style="background-color: #bbffbb; font-weight: bold; padding: 2px; border-radius: 4px;">{inserted_text}</span>')
            
        # DELETE: Model missed words (Red in Truth)
        elif opcode == 'delete':
            deleted_text = " ".join(a_words[a0:a1])
            html_truth.append(f'<span style="background-color: #ffcccc; text-decoration: line-through; padding: 2px; border-radius: 4px;">{deleted_text}</span>')
            
        # REPLACE: Mismatch (Red in Truth, Green in Pred)
        elif opcode == 'replace':
            deleted_text = " ".join(a_words[a0:a1])
            inserted_text = " ".join(b_words[b0:b1])
            html_truth.append(f'<span style="background-color: #ffcccc; text-decoration: line-through; padding: 2px; border-radius: 4px;">{deleted_text}</span>')
            html_pred.append(f'<span style="background-color: #bbffbb; font-weight: bold; padding: 2px; border-radius: 4px;">{inserted_text}</span>')
            
    return " ".join(html_truth), " ".join(html_pred)

print("✅ Highlighter function ready")

✅ Highlighter function ready

# Filter for errors (only works with old eval results format)
# Skip this section if using audit batch results
if results and 'match_type' in results[0]:
    errors = [r for r in results if r['match_type'] != 'exact']

    if errors:
        print(f"🔍 Analyzing {len(errors)} non-exact matches.")
        print(f"   Many of these are likely LABEL NOISE - the model correcting transcription errors!")
        
        # Start HTML Table
        # NOTE: CSS curly braces are doubled {{}} to escape them from Python string formatting
        html = """
        <style>
            .diff-table td {{ vertical-align: top; padding: 8px; border-bottom: 1px solid #ddd; }}
            .diff-table th {{ text-align: left; background-color: #f2f2f2; padding: 10px; }}
        </style>
        <h3>🖍️ Word-by-Word Diff Analysis</h3>
        <p><strong>Legend:</strong> 🔴 Red/Strikethrough = In ground truth but model missed | 🟢 Green/Bold = Model added or changed</p>
        <table class="diff-table" style='width:100%; border-collapse: collapse;'>
        <tr>
            <th style="width: 150px;">Play Audio</th>
            <th>Ground Truth (with diffs)</th>
            <th>Model Prediction (with diffs)</th>
        </tr>
        """
        
        for r in errors:
            # Create Audio Player
            try:
                with open(r['audio_path'], "rb") as f:
                    b64 = base64.b64encode(f.read()).decode()
                    audio_html = f'<audio controls style="width: 140px; height: 30px;"><source src="data:audio/wav;base64,{b64}" type="audio/wav"></audio>'
            except:
                audio_html = "🔇 Missing"

            # Generate Highlights
            hl_truth, hl_pred = highlight_differences(r['ground_truth'], r['prediction'])
            
            # Add Row
            html += f"<tr>"
            html += f"<td>{audio_html}<br><small style='color:grey'>{r['match_type'].upper()}</small><br><small style='color:grey'>{r['id']}</small></td>"
            html += f"<td style='font-family: monospace; font-size: 1.05em; line-height: 1.6;'>{hl_truth}</td>"
            html += f"<td style='font-family: monospace; font-size: 1.05em; line-height: 1.6;'>{hl_pred}</td>"
            html += "</tr>"
        
        html += "</table>"
        display(HTML(html))
    else:
        print("✅ No errors found! Model is perfect!")
else:
    print("ℹ️  Skipping old eval format analysis (use audit batch section below instead)")

ℹ️  Skipping old eval format analysis (use audit batch section below instead)

# Pattern analysis (only works with old eval results format)
if results and 'match_type' in results[0]:
    errors = [r for r in results if r['match_type'] != 'exact']
    
    if errors:
        from collections import Counter
        
        # Find words in ground truth but not prediction ("missed")
        missed_words = []
        added_words = []
        
        for e in errors:
            gt_words = set(e['ground_truth'].lower().split())
            pred_words = set(e['prediction'].lower().split())
            
            missed_words.extend(gt_words - pred_words)
            added_words.extend(pred_words - gt_words)
        
        print("🔍 Most commonly 'missed' words (often label noise):")
        for word, count in Counter(missed_words).most_common(10):
            print(f"   - '{word}': {count} times")
        
        print("\n🔍 Most commonly 'added' words (model corrections):")
        for word, count in Counter(added_words).most_common(10):
            print(f"   - '{word}': {count} times")
        
        print("\n💡 Interpretation:")
        print("   Articles like 'the', 'a', 'an' are often label noise.")
        print("   The model may be more faithful to the actual audio than the transcriber!")
else:
    print("ℹ️  Pattern analysis only available for old eval format.")

ℹ️  Pattern analysis only available for old eval format.

# Load audit batch results
AUDIT_FILE = "../output/audit_batch_results.json"

try:
    with open(AUDIT_FILE, 'r') as f:
        audit_results = json.load(f)
    print(f"✅ Loaded {len(audit_results)} audit samples")
    
    # Filter for disagreements only
    disagreements = [r for r in audit_results if r['is_disagreement']]
    print(f"🔍 Found {len(disagreements)} disagreements to verify ({len(disagreements)/len(audit_results)*100:.1f}%)")
    print(f"📊 Agreements: {len(audit_results) - len(disagreements)}")
    
except FileNotFoundError:
    print("❌ Audit batch file not found!")
    print("   Run: python scripts/generate_audit_batch.py")
    audit_results = []
    disagreements = []

✅ Loaded 100 audit samples
🔍 Found 63 disagreements to verify (63.0%)
📊 Agreements: 37

if disagreements:
    # Sort by WER (highest first) to prioritize major disagreements
    disagreements_sorted = sorted(disagreements, key=lambda x: x['wer'], reverse=True)
    
    # Build HTML table
    # NOTE: CSS curly braces are doubled {{}} to escape them from Python .format()
    html = """
    <style>
        .audit-table td {{ vertical-align: top; padding: 10px; border-bottom: 1px solid #ddd; }}
        .audit-table th {{ text-align: left; background-color: #f0f8ff; padding: 12px; font-weight: bold; }}
        .wer-badge {{ 
            background-color: #ff6b6b; 
            color: white; 
            padding: 3px 8px; 
            border-radius: 12px; 
            font-size: 0.85em; 
            font-weight: bold;
        }}
        .note-box {{
            width: 100%;
            min-height: 40px;
            border: 1px solid #ccc;
            padding: 5px;
            font-family: monospace;
            font-size: 0.9em;
        }}
    </style>
    <h3>🎧 Disagreement Analysis ({} samples)</h3>
    <p><strong>Legend:</strong> 🔴 Red/Strikethrough = In ground truth but model missed | 🟢 Green/Bold = Model added or changed</p>
    <table class="audit-table" style='width:100%; border-collapse: collapse;'>
    <tr>
        <th style="width: 50px;">#</th>
        <th style="width: 180px;">Audio & WER</th>
        <th style="width: 40%;">Ground Truth (Label)</th>
        <th style="width: 40%;">Model Prediction</th>
    </tr>
    """.format(len(disagreements_sorted))
    
    for idx, r in enumerate(disagreements_sorted, 1):
        # Create Audio Player
        try:
            with open(r['audio_path'], "rb") as f:
                b64 = base64.b64encode(f.read()).decode()
                audio_html = f'<audio controls style="width: 160px;"><source src="data:audio/wav;base64,{b64}" type="audio/wav"></audio>'
        except Exception as e:
            audio_html = f"🔇 <small>Error: {str(e)[:20]}</small>"

        # Generate Diff Highlights
        hl_truth, hl_pred = highlight_differences(r['ground_truth'], r['prediction'])
        
        # WER Badge
        wer_pct = r['wer'] * 100
        wer_badge = f'<span class="wer-badge">{wer_pct:.1f}% WER</span>'
        
        # Build Row
        html += f"<tr>"
        html += f"<td style='text-align: center; font-weight: bold; color: #666;'>{idx}</td>"
        html += f"<td>{audio_html}<br><br>{wer_badge}<br><small style='color:grey; font-size: 0.8em;'>{r['id']}</small></td>"
        html += f"<td style='font-family: monospace; font-size: 1em; line-height: 1.8; padding: 10px; background-color: #fff5f5;'>{hl_truth}</td>"
        html += f"<td style='font-family: monospace; font-size: 1em; line-height: 1.8; padding: 10px; background-color: #f0fff4;'>{hl_pred}</td>"
        html += "</tr>"
    
    html += "</table>"
    display(HTML(html))
    
    print("\n" + "="*80)
    print("📝 MANUAL VERIFICATION INSTRUCTIONS")
    print("="*80)
    print("Listen to each audio clip and mark your findings:")
    print("- If MODEL IS WRONG → Count as 'Model Error'")
    print("- If MODEL IS CORRECT (label is wrong) → Count as 'Label Noise'")
    print("\nAfter listening to all disagreements, use the cell below to calculate noise rate.")
    
else:
    print("✅ No disagreements found - perfect match!")

================================================================================
📝 MANUAL VERIFICATION INSTRUCTIONS
================================================================================
Listen to each audio clip and mark your findings:
- If MODEL IS WRONG → Count as 'Model Error'
- If MODEL IS CORRECT (label is wrong) → Count as 'Label Noise'

After listening to all disagreements, use the cell below to calculate noise rate.

# ============================================
# 📝 MANUAL INPUT REQUIRED
# ============================================
# After listening to all disagreements above, enter your counts here:

# --- 📝 AUDIT TRACKER (POPULATED) ---

# 1. LABEL NOISE (The Model was RIGHT / GT was WRONG)
# Includes text misses (1,2,3...), name fixes (4,8), and grammar fixes
label_error_ids = [
    1, 2, 3, 4, 5, 6, 8, 9, 11, 12, 13, 16, 17, 22, 23, 
    25, 27, 28, 29, 32, 33, 37, 39, 40, 42, 43, 44, 46, 
    48, 49, 52, 53, 56, 57, 59, 62
]

# 2. MODEL ERRORS (The Ground Truth was RIGHT / Model was WRONG)
# Includes accents (51), hallucinations (38), and misspellings
model_error_ids = [
    14, 19, 21, 24, 26, 31, 34, 38, 41, 47, 51, 55, 61, 63
]

# 3. BOTH WRONG / HARD (Accents, ambiguous audio)
# Includes "selfishly" (45) and other complex cases
ambiguous_ids = [
    10, 15, 18, 20, 30, 35, 36, 45, 50, 54, 58
]

# 4. NORMALIZATION (Harmless differences)
normalization_ids = [
    7  # Just a period
]

# --- AUTOMATIC CALCULATOR ---
total_audited = 100
noise_count = len(label_error_ids)
model_fail_count = len(model_error_ids)
ambiguous_count = len(ambiguous_ids)
norm_count = len(normalization_ids)

print(f"📊 FINAL AUDIT REPORT")
print(f"=====================")
print(f"Total Samples Audited: {total_audited}")
print(f"---------------------")
print(f"✅ Label Noise (Model Won): {noise_count} ({noise_count/total_audited*100:.1f}%)")
print(f"❌ Model Errors:            {model_fail_count} ({model_fail_count/total_audited*100:.1f}%)")
print(f"⚠️  Both/Ambiguous:          {ambiguous_count} ({ambiguous_count/total_audited*100:.1f}%)")
print(f"ℹ️  Normalization:           {norm_count} ({norm_count/total_audited*100:.1f}%)")
print(f"---------------------")
print(f"🔍 CONCLUSION: In {noise_count}% of cases, the model outperformed the ground truth.")

📊 FINAL AUDIT REPORT
=====================
Total Samples Audited: 100
---------------------
✅ Label Noise (Model Won): 36 (36.0%)
❌ Model Errors:            14 (14.0%)
⚠️  Both/Ambiguous:          11 (11.0%)
ℹ️  Normalization:           1 (1.0%)
---------------------
🔍 CONCLUSION: In 36% of cases, the model outperformed the ground truth.

python scripts/generate_audit_batch.py

🖍️ Diff-Highlighted Error Analysis¶

Key Finding: Label Noise Detected!¶

Load Old Eval Results (Optional)¶

The Highlighter Function 🖍️¶

🎧 Interactive Dashboard with Diff Highlighting¶

📊 Pattern Analysis¶

🎯 Key Insights¶

🔬 Label Noise Audit (100-Sample Batch)¶

🎧 Interactive Verification Interface¶

🎧 Disagreement Analysis (63 samples)

📊 Calculate Label Noise Rate¶

📋 Audit Workflow & Findings¶

#	Audio & WER	Ground Truth (Label)	Model Prediction
1	120.0% WER test_70	HAVE A LOOK AT CHINA AND SAUDI ARABIA THEY FILTER CONTENT ACCORDING TO POLITICAL IDEAS	if you have a look at china and saudi arabia they fill the content according to political ideas
2	35.3% WER test_29	AND TWENTY THE GREAT INNOVATION UNION DIGITAL ACCESS TO ALL THE NEXT GENERATION NETWORK AND SO FORTH	THE GREAT INNOVATION UNION THE DIGITAL ACCESS TO ALL THE NEXT GENERATION NETWORK AND SO ON AND SO FORTH
3	28.6% WER test_48	YOU CAN DECLARE ADOPTED OR NOT ADOPTED	SO YOU CAN DECLARE IT ADOPTED OR NOT ADOPTED
4	26.2% WER test_152	MR EFOVI IF YOU WANT TO FLY WITH YOUR OR OUR AIRBUS THREE HUNDRED AND EIGHTY ENERGY PACKAGE YOU MUST DISTRIBUTE AS A GOOD MANAGER RESPONSIBILITIES TO EVERY MEMBER OF THE CREW RESPONSIBILITIES WITH HIGHER PRIORITY THAN FLEXIBILITY IN EACH MEMBER STATE	IF MR ŞEFÇOVIČ YOU WANT TO FLY WITH YOUR OR OUR AIRBUS A380 ENERGY PACKAGE YOU MUST DISTRIBUTED AS A GOOD MANAGER THE RESPONSIBILITIES TO EVERY MEMBER OF THE CREW RESPONSIBILITIES WITH HIGHER PRIORITY THAN THE FLEXIBILITY OF EACH MEMBER STATE
5	22.2% WER test_39	EFFORT IF WE DO NOT DO THAT WE WILL	IF WE DO NOT DO THAT WE WILL LOSE
6	20.0% WER test_134	TODAY EUROPE IS FREE AND REUNITED AND I WANT TO THANK THIS HOUSE AND EACH AND EVERY ONE WHO DARED TO SPEAK OUT AT THAT TIME FOR TRUTH AND FREEDOM	TODAY EUROPE IS FREE AND REUNITED AND I WANT TO THANK THIS HOUSE AND EACH AND EVERYONE WHO DARED TO SPEAK AT THAT TIME TO SPEAK OUT FOR TRUTH AND FREEDOM
7	20.0% WER test_98	THEY CANNOT GO TO SCHOOL	THEY CANNOT GO TO SCHOOL.
8	19.1% WER test_56	THAT HARSH TREATMENT WAS METED OUT BY MR PTTERING AND MR SIWIEC THE VICE PRESIDENT REPLACING HIM LATER IN THE AFTERNOON	THAT HARSH TREATMENT WAS METED OUT BY MR PUTTERING AND MR MAJKECIEVIC THE VICEPRESIDENT REPLACING HIM LATER IN THE AFTERNOON
9	18.2% WER test_185	THEREFORE WE WOULD LIKE TO PROTEST AGAINST CHANGING THE LEGAL BASIS	AND THEREFORE WE WOULD LIKE TO PROTEST AGAINST CHANGING THE LEGAL BASE
10	16.7% WER test_87	I DO NOT WANT TO DEPICT A DOOM SCENARIO FOR THE FUTURE NOR DO I WANT TO LOOK BACK IN ANGER ABOUT THE FAILURE OF COPENHAGEN ALTHOUGH I AM ANGRY THEREFORE THE RESOLUTION IS TO DO FAR BETTER IN THE FUTURE THE NEXT OPPORTUNITY BEING MEXICO THIS YEAR	I DO NOT WANT TO DEPICT A DOOMSCENARIO FOR THE FUTURE NOR DO I WANT TO LOOK BACK IN ANGER ABOUT THE FAILURE OF COPENHAGEN ALTHOUGH I AM ANGRY THEREFORE THREE SOLUTIONS TO DO FAR BETTER IN THE FUTURE THE NEXT STEP AT DUTYUNITY BEING MEXICO THIS YEAR
11	16.0% WER test_116	IT IS NOT ONLY IN EUROPE IT IS ALL OVER THE WORLD AND WE HAVE A RESPONSIBILITY TO SHOW THE WAY AND LEAD THE WAY	IT'S NOT ONLY IN EUROPE IT'S ALL OVER THE WORLD AND WE HAVE A RESPONSIBILITY TO SHOW THE WAY AND LEAD THE WAY
12	15.8% WER test_93	THE ARTICLE LIMITS EXISTING RIGHTS AND ELIMINATES WELLFUNCTIONING MINORITY LANGUAGE SCHOOL SYSTEMS WHICH HAVE WORKED VERY WELL SO FAR	THE ARTICLE LIMITS EXISTING RIGHTS AND ELIMINATES WELL FUNCTIONING MINORITY LANGUAGE SCHOOL SYSTEMS WHICH WORKED VERY WELL SO FAR
13	15.4% WER test_189	I KNOW IT MEANS AS MUCH TO THEM AS IT MEANS TO ME	SO IT MEANS AS MUCH TO THEM AS IT MEANS TO ME
14	15.0% WER test_166	ARE YOU WILLING TO ACT IN FAVOUR OF THE SOCIAL DIMENSION TO BE INCLUDED IN THE EU COMPETENCIES AS PROPOSED	ARE YOU WILLING TO ACT IN FAVOUR OF THE SOCIAL DIMENSION BEING INCLUDED IN THE EU COMPETENCES AS PROPOSED
15	13.5% WER test_55	TO AVOID ANY SUSPICION THAT THE COUNCIL IN THIS SITUATION WOULD TAKE THE ADOPTION OF AMENDING BUDGET NO SIX AS AN ARGUMENT FOR DELAYING AND NOT ADOPTING AMENDING BUDGET NO EIGHT MY GROUP HAS TABLED AN AMENDMENT IN ORDER TO LINK THE ADOPTION OF AMENDING BUDGET NO SIX WITH AMENDING BUDGET NO	TO AVOID ANY SUSPICION THAT THE COUNCIL IN THIS SITUATION WOULD TAKE THE ADOPTION OF AMENDING BUDGETARY RULES AS AN ARGUMENT FOR DELAYING OR NOT ADOPTING AMENDING BUDGET EIGHT MY GROUP HAS TABLED AN AMENDMENT IN ORDER TO LINK THE ADOPTION OF AMENDING BUDGET RULE SIX WITH AMENDING BUDGET EIGHT
16	13.3% WER test_164	BUT AS A SOCIALIST OF COURSE IT IS VERY EASY TO SPEND OTHER PEOPLE'S MONEY	BUT AS A SOCIALIST OF COURSE IT'S VERY EASY TO SPEND OTHER PEOPLE'S MONEY
17	12.5% WER test_6	SECONDLY THE COURT HAS BEEN ADAMANT ABOUT HIGHLIGHTING THE IMPORTANCE OF THE FULL COMMITMENT OF MEMBER STATES IN ENSURING BETTER RULES AND BETTER SPENDING	SECOND THE COURT HAS BEEN ADAMANT ON HIGHLIGHTING THE IMPORTANCE OF FULL COMMITMENT OF MEMBER STATES IN ENSURING BETTER RULES AND BETTER SPENDING
18	12.5% WER test_71	THE ECONOMIC BURDEN OF THESE DISEASES IS PUTTING PRESSURE ON THE MEMBER STATES AND THE COSTS SIGNIFICANTLY INCREASE WITH THE PROGRESSION OF THE DISEASES	THE ECONOMIC BURDEN OF THESE DISEASES IS PUTTING PRESSURE ON THE MEMBER STATES AND THE COST SIGNIFICANTLY INCREASES WITH THE PROGRESSION OF THE DISEASE
19	12.5% WER test_199	A GOVERNMENT THAT HAS SHOWN ITS DISRESPECT FOR MOST OF OUR VALUES FOR ALMOST FOUR DECADES	A GOVERNMENT THAT HAS SHOWN ITS DISREGARD FOR THE MOST OF OUR VALUES FOR ALMOST FOUR DECADES
20	12.0% WER test_137	WE ARE SENDING THE MESSAGE THAT A SOCIETY CAN ONLY HAVE A HEALTHY ECONOMY WHEN ITS MEMBERS ARE ABLE TO CONTRIBUTE FULLY TO ITS DEVELOPMENT	WE ARE SENDING THE MESSAGE THAT SOCIETY CAN HAVE A HEALTHY ECONOMY ONLY WHEN ITS MEMBERS ARE ABLE TO CONTRIBUTE FULLY TO ITS DEVELOPMENT
21	11.8% WER test_26	HOW ARE WE GOING TO MEASURE WHETHER THE INFLUX IS HIGH NOT HIGH OR HIGH ENOUGH WHEN IT IS ALL OVER THE EUROPEAN UNION HAS TO DECIDE WHETHER IT WANTS TO ACT OR REACT	HOW ARE WE GOING TO MEASURE WHETHER THE INFLATION IS HIGH NOT HIGH HIGH ENOUGH WHEN IT'S ALL OVER THE EUROPEAN UNION HAS TO DECIDE WHETHER IT WANTS TO ACT OR REACT
22	11.8% WER test_82	I AGREE WITH THE INTENTION TO ENSURE THAT END USERS WILL BE ABLE TO RECEIVE FULL INFORMATION ON THE LABEL EVEN IF THE PRODUCT IS BOUGHT AT A DISTANCE VIA THE INTERNET OR TELEMARKETING	I AGREE WITH THE INTENTION TO ENSURE THAT END USERS WILL BE ABLE TO RECEIVE THE FULL INFORMATION OF THE LABEL EVEN IF THE PRODUCT IS BOUGHT BY DISTANCE VIA THE INTERNET OR TELEMARKETING
23	11.6% WER test_129	ENDING UNJUSTIFIED GEO BLOCKING PRACTICES IS ONE CONCRETE STEP IN THE RIGHT DIRECTION BUT I BELIEVE IT SHOULD BE DONE IN A WAY THAT DOES NOT HAMPER SMES AND START UPS AND DOES NOT RAISE PRICES FOR CONSUMERS ESPECIALLY IN NEWER MEMBER STATES	ENDING UNJUSTIFIED GEO BLOCKING PRACTICES IS ONE CONCRETE STEP IN THE RIGHT DIRECTION BUT I BELIEVE IT SHOULD BE DONE IN A WAY THAT DOES NOT HAMPER SMES AND STARTUPS AS WELL AS DOES NOT RAISE PRICES FOR CONSUMERS ESPECIALLY NEWER MEMBER STATES
24	11.6% WER test_197	THIRDLY THE COMMISSION DELEGATED REGULATION OF THIRTY SEPTEMBER TWO THOUSAND AND THIRTEEN ON THE MODEL FINANCIAL REGULATION FOR PUBLIC PRIVATE PARTNERSHIP BODIES WILL ENTER INTO FORCE IN ORDER TO ALLOW JOINT UNDERTAKINGS TO BENEFIT FROM THE SIMPLIFICATIONS INTRODUCED IN THE NEW FINANCIAL FRAMEWORK	THREE COMMISSION'S DELEGATED REGULATION OF THIRTY SEPTEMBER TWO THOUSAND AND THIRTEEN ON THE MODEL FINANCIAL REGULATION FOR THE PUBLIC PRIVATE PARTNERSHIP BODIES WILL ENTER INTO FORCE IN ORDER TO ALLOW THE JOINT UNDERTAKINGS TO BENEFIT FROM THE SIMPLIFICATIONS INTRODUCED IN THE NEW FINANCIAL FRAMEWORK
25	11.1% WER test_75	TWO THOUSAND AND SEVENTEEN WAS THE YEAR WHEN WE SAW OBSTACLES TO RESOLVABILITY OF THE VENETIAN BANKS AND THE FIRST RESOLUTION UNDER THE EU FRAMEWORK THE POPULAR CASE WHICH SHOWS THAT FURTHER TRANSPARENCY IS CLEARLY NEEDED	TWO THOUSAND AND SEVENTEEN WAS THE YEAR WHEN WE SAW THE OBSTACLES TO RESOLVABILITY OF THE VENETIAN BANKS AND THE FIRST RESOLUTION UNDER THE EU FRAMEWORK THE POPULAR CASE HAS SHOWN FURTHER TRANSPARENCY IS CLEARLY NEEDED
26	10.0% WER test_108	FOUR HUNDRED AND FIFTY THREE TO FOUR FIVE HUNDRED AND FORTY SIX UNFORTUNATELY A VERY HIGH NUMBER OF HEALTH PROFESSIONALS ARE AFFECTED WITH FOUR HUNDRED AND TWENTY SEVEN DOCTORS AND NURSES SICK AND OF THOSE TWO HUNDRED AND THIRTY HAVE LOST THEIR LIVES TRYING TO SAVE THE LIVES OF OTHERS	FOUR HUNDRED AND FIFTY THREE TO FOUR HUNDRED AND FIVE HUNDRED AND FORTY SIX AND UNFORTUNATELY A VERY HIGH NUMBER OF HEALTH PROFESSIONALS ARE AFFECTED FOUR HUNDRED AND TWENTY SEVEN DOCTORS AND NURSES SICK AND OF THOSE TWO HUNDRED AND THIRTY LOST THEIR LIVES TRYING TO SAVE THE LIVES OF OTHERS
27	10.0% WER test_147	BY WORKING TOGETHER BY ACTING TOGETHER WE DEFINE WHO WE	BY WORKING TOGETHER BY ACTING TOGETHER WE DEFINE WHO WE ARE
28	10.0% WER test_198	PARLIAMENT IS ALSO APPEALING TO NATIONAL LAWMAKERS TO DISTINGUISH CLEARLY HIGHER RISK OR LOWER LIQUIDITY ASSETS FROM THOSE ASSETS WHICH ARE ELIGIBLE FOR UCITS TYPE COVERED BONDS LEAVING SME CREDITS INFRASTRUCTURE INVESTMENTS AND CONSUMER CREDITS TO A NEW INSTRUMENT WHICH AS I HAVE SAID WOULD BE CALLED EUROPEAN SECURED NOTES	PARLIAMENT ALSO APPEALS TO NATIONAL LAWMAKERS TO CLEARLY DISTINGUISH HIGHER RISK OR LOWER LIQUIDITY ASSETS FROM THOSE ASSETS WHICH ARE ELIGIBLE FOR USES TYPE COVERED BONDS LEAVING SME CREDITS INFRASTRUCTURE INVESTMENTS AND CONSUMER CREDITS TO A NEW INSTRUMENT WHICH AS I HAVE SAID WOULD BE CALLED EUROPEAN SECURED NOTES
29	9.5% WER test_188	EVEN THE UK GOVERNMENT'S OWN SO CALLED BALANCE OF COMPETENCES REVIEW SHOWS THAT FOREIGN POLICY COMPETENCES REMAIN SQUARELY WITH THE MEMBER STATES AND THAT MOST OF THE EVIDENCE ARGUES STRONGLY THAT IT IS IN THE UK'S INTEREST TO WORK THROUGH THE EU	EVEN THE UK GOVERNMENT'S OWN SO CALLED BALANCE OF COMPETENCES REVIEW SHOWS THAT FOREIGN POLICY COMPETENCES REMAIN SQUARELY WITH MEMBER STATES AND THAT MOST OF THE EVIDENCE ARGUES STRONGLY IN THE UK'S INTEREST TO WORK THROUGH THE EU
30	8.8% WER test_22	IF THE COMMISSION'S PLAN IS AN EXAMPLE FOR THE REST OF THE MEDITERRANEAN AND THEY LOBBY FOR IT THE WAY THEY LOBBIED FOR THIS REPORT THEN I DON'T KNOW WHY WE ALL ARE HERE	IF THE COMMISSION'S PLAN IS AN EXAMPLE FOR THE REST OF THE MEDITERRANEAN AND THEY LOBBIED FOR IT THE WAY THEY LOBBIED FOR THIS REPORT THEN I DON'T KNOW WHY WE ARE ALL HERE
31	8.7% WER test_4	AND COULD YOU PLEASE ALSO TELL ME WHAT IN LONDON YOU ARE SUPPORTING AS MEASURES IN THE CITY AGAINST THE INTERNATIONAL MONEYLAUNDERING SYSTEMS	AND COULD YOU PLEASE ALSO TELL ME WHAT IN LONDON YOU ARE SUPPORTING AS MEASURES IN THE CITY AGAINST THE INTERNATIONAL MONEY LAUNDERING SYSTEMS
32	8.3% WER test_40	MS RA THUN FOR HER COMMITMENT AND A GREAT JOB DURING THE NEGOTIATION PROCESS WITH THE COMMISSION AND THE COUNCIL DURING THE ESTONIAN PRESIDENCY	MS ROSA THUN FOR HER COMMITMENT AND A GREAT JOB DURING THE NEGOTIATION PROCESS WITH THE COMMISSION AND A COUNCIL DURING THE ESTONIAN PRESIDENCY
33	8.3% WER test_88	THE THRUST OF THIS DISCHARGE REPORT IS CRYSTAL CLEAR THE ECONOMIC AND FINANCIAL CRISIS HAS GREATLY INCREASED THE DEMAND FOR HIGH QUALITY PUBLIC SPENDING	THE THRUST OF THIS DISCHARGED REPORT IS CRYSTAL CLEAR THE ECONOMIC AND FINANCIAL CRISIS HAS GREATLY RISEN THE DEMAND FOR HIGH QUALITY PUBLIC SPENDING
34	8.3% WER test_117	WE SHOULD CONTINUE WITH THE EFFORTS TO INVOLVE THOSE COUNTRIES MORE INTIMATELY	WE SHOULD CONTINUE WITH OUR EFFORTS TO INVOLVE THOSE COUNTRIES MORE INTIMATELY
35	8.0% WER test_62	AT THE SAME TIME THE NEW DIRECTIVE WILL PROVIDE A MINIMUM LEVEL OF PROTECTION FOR LINKED TRAVEL ARRANGEMENTS WHICH ARE LOOSER COMBINATIONS OF TRAVEL SERVICES	AT THE SAME TIME THE NEW DIRECTIVE WOULD PROVIDE A MINIMUM LEVEL OF PROTECTION FOR LINKED TRAVEL ARRANGEMENTS WHICH ARE LOSER COMBINATIONS OF TRAVEL SERVICES
36	7.7% WER test_9	AND NATIONAL COMPETITION AUTHORITIES BUT THE REPORT ALSO SAYS VERY CLEARLY THAT THIS INDEPENDENCE IS STRONGLY LINKED TO THE AVAILABILITY OF HUMAN AND FINANCIAL RESOURCES HOWEVER	AND NATIONAL COMPETITION AUTHORITIES BUT ALSO THE REPORT SAYS VERY CLEARLY THAT THIS INDEPENDENCE IS STRONGLY LINKED TO THE AVAILABILITY OF HUMAN AND FINANCIAL RESOURCES HOWEVER
37	7.7% WER test_99	THE SERVICES MUST NOT STOP AT THE INTERNAL BORDERS OF THE EUROPEAN UNION	THE SERVICES MUST NOT STOP ON THE INTERNAL BORDERS OF THE EUROPEAN UNION
38	7.4% WER test_150	I WANTED TO PAY TRIBUTE TO THE MALTESE GOVERNMENT AND TO THE PRIME MINISTER I WANT TO PAY TRIBUTE TO WHAT THE PRIME MINISTER OF MALTA DID	I WANTED TO PAY TRIBUTE TO THE MALTESE GOVERNMENT AND TO THE PRIME MINISTER OF MALTA. I WANT TO PAY TRIBUTE TO WHAT THE PRIME MINISTER OF MALTA DID
39	6.7% WER test_28	WE ARE CONVINCED THAT IT'S VERY USEFUL FOR THE QUALITY OF THE FUNDING PROGRAMMES IT IS BASED ON A NEW FOCUS ON RESULTS AND MILESTONES THAT HAVE TO BE ACHIEVED	WE ARE CONVINCED THAT IT IS VERY USEFUL FOR THE QUALITY OF THE FUNDING PROGRAMMES IT IS BASED ON A NEW FOCUS ON RESULTS AND MILESTONES THAT HAVE TO BE ACHIEVED
40	6.7% WER test_180	VICE PRESIDENT TAJANI HAS STATED THAT INDUSTRY IS AT THE HEART OF EUROPE AND IS INDISPENSABLE FOR FINDING SOLUTIONS TO THE CHALLENGES OF OUR SOCIETY TODAY AND IN THE FUTURE	VICE PRESIDENT AYANIS STATED THAT INDUSTRY IS AT THE HEART OF EUROPE AND IS INDISPENSABLE FOR FINDING SOLUTIONS TO THE CHALLENGES OF OUR SOCIETY TODAY AND IN THE FUTURE